You can't prove an AI-assisted review is accurate. You can make it checkable.

Part 1 of 3 in AI-assisted desk research. Part 2: How to do an AI-assisted literature review. Part 3, a kit of prompts for it, is in testing.

I was doing research where a true systematic review wasn’t going to work. The peer-reviewed evidence was too thin. So I ran a multivocal review, which pulls in practitioner and grey sources next to the papers. Partway through I caught myself asking a question I’d been skipping: how do I know I can trust what the AI is handing back?

My answer is that you can’t prove an AI-assisted review is accurate. Accuracy depends on the task, the model, the run, and the date, and the numbers in the papers don’t transfer to your topic. What you can do is make the review checkable. Every claim traces to a passage, table, or figure you read in the source. That’s a lower bar than “accurate,” and it’s one you can actually clear.

A quick primer if you’re new to this. A systematic literature review (SLR) is a review done by a defined, structured method. In practice it typically covers the formally published literature, such as journal and conference papers. A multivocal literature review (MLR) adds grey literature, such as blog posts, videos, and white papers. The standard MLR guidelines put it this way: an MLR is “a union of the sources that would be studied in an SLR and in a GLR”. Reviewers reach for one when “the topic is emerging, when practice is evolving faster than scholarly publications, or when practitioner perspectives are central”. That was my situation.

Why the model’s accuracy can’t be your guarantee

Start with what the failures look like. In a 2024 test, GPT-3.5 and GPT-4 were asked to find studies for eleven published reviews. Of the references they returned, 39.6% (GPT-3.5) and 28.6% (GPT-4) were “hallucinated”, which the authors defined as at least two of the title, first author, and year being wrong. In that test, the models retrieved roughly one in eight of the real studies. A bad reference looks like a real one. I’ve seen it myself. In 2023, I asked ChatGPT for a citable reference for a claim. It gave me an author, a title, a publication date, and quotes from the PDF. I couldn’t find the paper anywhere online, so I reached out to the author. He told me he had never written it. I suspect the model attached his name because he was a domain expert. Anyone who accepted that citation because it seemed reasonable would have been wrong.

Extraction errors are harder to spot, because a wrong number sits inside a real paper. One small proof-of-concept study had Claude 2 pull 160 data points from ten trial reports. It got 154 right. Of the six errors, one was a wrong effect estimate for the primary outcome, and the authors describe invented exclusion criteria as hallucination. Another study, using GPT-4 on ten papers, got study settings right 28 times out of 30 and behavioral components right 15 times out of 30. The more the answer depends on reading between the lines, the worse it tends to get.

Then there’s the part that surprised me. On screening, the evidence is more encouraging. A systematic review of 63 LLM studies found a median agreement with human reviewers of 0.92 on relevant abstracts and 0.89 on irrelevant ones. The same review found extraction accuracy ranging from 0.36 to 1.00, and a median of 0.62 for risk of bias assessment (only six assessments, so hold that loosely). Search looks worst. A scoping review of 37 studies found it was the step most often rated not promising, and a medical library’s guide rates AI Low for searching, quality assessment, and data collection. So “how accurate is AI at reviews?” has no single answer. It depends on the task and the setup.

It also has one answer per run. A study that set five models to “deterministic” settings still saw accuracy swing by up to 15% across runs. That wasn’t a literature review study, so I wouldn’t borrow the size of the swing. But an editorial in Research Synthesis Methods says the same thing in plainer terms: outputs “can vary depending on the prompt, the model version, and random chance”. And most of it is from 2023 and 2024. Nine in ten of the 63 studies were published in those two years. The scoping review’s search ended in February 2024. The 63-study review’s ended in January 2025.

One more wrinkle. The yardstick is people, and people miss things too. The authors of the 63-study review report agreement and not sensitivity, because human reviewers aren’t a reference standard. In a randomized trial with 280 screeners, a single reviewer missed 13% of relevant studies and two reviewers missed 3%. So “as good as a human” is a moving target.

Now the strongest objection: screening is solved enough, so this is overkill. I’d concede most of it. At scale, one study of model ensembles tested six models over 119,695 abstracts from Cochrane reviews. With each model’s best prompt, sensitivity ranged from 0.756 to 1.000. A GPT-4 study’s authors wrote that for certain review tasks under specific conditions, LLMs “can rival human performance”. But look at the price. In that study, precision with those same prompts fell to between 0.004 and 0.096, which the authors attribute to how rare relevant records are. Rare or not, that’s a lot of irrelevant records to read. A dual-model preprint hit 99.7% sensitivity with 49.3% specificity, which means about half the records that were later excluded stayed in the pile. The authors call that figure conservative, because the models saw only abstracts while the reference standard came from full-text decisions. And screening is one step. Those benchmarks score screening decisions, not the sentence you eventually write in your findings.

The version I’d trust keeps a person in charge. In one small pilot, four neurology trainees screened 1,003 articles. The two who got an LLM-written summary beside each abstract screened about four times faster, with 100% sensitivity against 88% and 92% for the two who didn’t. Each trainee tried only one condition, and it’s four people in one specialty, so it’s a pointer, not proof. It does match what the 63-study authors wrote: LLMs “should not be used because they are novel but because they are validated in this context.”

Check the claim, not the model

If accuracy can’t be your guarantee, something else has to be. The four big evidence synthesis bodies (Cochrane, Campbell, JBI, and the Collaboration for Environmental Evidence) agree in their 2025 position statement. They write: “AI and automation in evidence synthesis should be used with human oversight.” Anything that makes or suggests a judgement gets fully reported, and synthesists stay accountable. That’s what “checkable” means in practice. A person answers for each claim, and a person can see how it got there. But the statement sets out principles and points to the RAISE recommendations for more detail. Proposed checklists are starting to appear, such as ARISMA for AI-assisted systematic and scoping reviews and PRISMA-trAIce for reporting AI use. Both are proposals, not settled standards. I did not find one written for everyday desk research. This article is my attempt at that.

The second objection is that the guidelines already cover this. Mostly, they do. PRISMA 2020 already asks you to report the automation tools you used in selection. The MLR guidelines already give you a decision aid, quality criteria, and stopping rules. The wording of your inclusion criteria moves model performance, which is one more reason to write the protocol before you open a chat window. And that editorial tells anyone testing AI in reviews to document how the prompt was developed, and to report the API version and access date for commercial models. I’d copy that for anyone, not just testers. The 63-study authors found prompt details and model versions scarcely reported. In a separate review of LLM use in reviews, only 15.1% of the papers were “actual reviews that acknowledged LLM usage”.

The method here is the old one, run on purpose, with three additions.

Three additions

Capture the exact quote when you retrieve it. When the model hands me a claim, I save the sentence it came from, word for word, at that moment. A paraphrase is hard to check, because you can’t tell which sentence it came from. A quote that passes through a fetch tool can drift from the page. In my own run, one did, and seven could only be confirmed by fetching the page again. I label every quote by how I got it: copied from the page, pasted by me, or paraphrase. Paraphrases never go in the write-up inside quotation marks.

Audit every claim against the raw text. Each claim in the draft gets checked against the source it cites. My rule is simple: if a claim can’t point to a passage, table, or figure I’ve read, it doesn’t go in. That includes the AI’s helpful summary of a source. I read the source.

Get a second auditor, then audit the auditor. I hand the draft and sources to a separate AI and ask it to find what’s wrong. Then I check what it flagged, because it’s wrong sometimes too.

I’m not claiming the whole workflow is fast. One measurement I found is a study of six reviews where Claude drafted the data extraction and a person verified it. That cut extraction time by a median of 41 minutes per study, and accuracy was 91.0% against 89.0% for people working alone. But it checked extracted data against the PDFs, not claims in a write-up, so I can’t tell you what my full workflow costs. It spends some of what the AI saved, and that’s the trade.

In one run, I used this process on a research brief about building a program in a regulated area. The first audit checked 54 quotes. Forty-five matched the source exactly. Seven I confirmed against the fetched page. One matched once I accounted for a page break mid-sentence. One I corrected. The bigger catch wasn’t a quote. A research step had called a source “peer-reviewed,” and that label rode into the vetted summary. When I opened the PDF, it was conference proceedings with no peer-review statement. I changed the wording to say what the source was.

Then an external AI audit came back with 56 rows. Twenty-six held up, 23 were debatable, five were missing a citation, and two were marked wrong. Of those two, one was the auditor’s mistake. It marked my reference to a federal rule as wrong, and the agency’s own page confirmed the rule. The other was fair. I’d called a settlement “court-ordered” when “court-enforceable” was the accurate term. A second external audit caught a claim where my draft overstated what a government report found, and that one was right too. After the edits, I re-checked all 81 quotes. That’s one run, graded by me. It shows what the checks can catch. It doesn’t show the checks work in general.

I saw the same failure in my own setup for this article. When I ran Google’s AI search to seed the sources, it returned ten sources as titles and blurbs with redirect links. I searched for each by title. Three I never turned up, and one page I did find didn’t say what its blurb claimed. A source list that sounds right and can’t be resolved is close kin to a made-up reference, because you can’t check either one.

When the peer-reviewed evidence runs thin

That’s where an MLR earns its place. The guidelines give you a decision aid of seven questions, and one or more yes answers is the signal to include grey literature. The catch is that grey sources often rest on experience and opinion. So they get graded, and the guidelines are blunt about it: “There is no one-size-fits-all quality model for all types of GL.”

A recent MLR experience report shows one way to do it. The authors scored sources with two instruments by source type, normalized to 0 to 1, and dropped anything below 0.5. They replaced preprints with the published version when one existed. Their stop rule was the top 100 Google results plus the top 50 each from LinkedIn, Medium, and Reddit, continuing until a page had fewer than 50% relevant hits. They also said what you give up: “the transient nature of most grey sources, together with search-engine dynamics, compromises full replicability.”

I found only one test of an LLM on grey literature, and its authors call it the first. It used GPT-4 on a review that included grey sources. Its authors still wrote that “substantial caution should be exercised” when LLMs are used to conduct systematic reviews. So label every grey source as grey, grade it, and say in the write-up how.

What you get and what you don’t

Followed carefully, the workflow gets you one thing: each claim in your review points to text you read. Matching a sentence still isn’t the same as being right about it. It doesn’t guarantee you found everything. Remember that the models found only one in eight real studies when asked to search. Screening sensitivity ran from 0.756 to 1.000 depending on setup. It doesn’t guarantee the sources are right, which is why grey sources get graded. And it hasn’t been independently validated. It’s one practitioner’s workflow, built on one case.

Accuracy still matters, because it affects how much you’ll have to fix. It just can’t be your guarantee. Don’t ask whether the model is accurate. Ask whether you can check every claim.